Latest / Minds, Bodies, and Terawatts / Humanoid Robots Beat Human World Record in Beijing Half-Marathon

Transcript

Humanoid Robots Beat Human World Record in Beijing Half-Marathon · Minds, Bodies, and Terawatts

• 0:00
  [Host] This is Minds, Bodies, and Terawatts, where the finish line keeps moving. AP News reported on April 19th that a Chinese-made humanoid robot won the Beijing half-marathon, finishing faster than the human world record for the distance.
• 0:15
  [Guest] Let me say that again. A bipedal machine, built in China, covered 13.1 miles on two legs faster than any human being ever has. And about 40 percent of the robots in the race navigated the course autonomously, no remote operator steering them.
• 0:32
  [Host] Forty percent autonomous. Wasn't last year's race basically all remote-controlled?
• 0:36
  [Guest] The inaugural Beijing robot half-marathon was in 2026, and autonomy was negligible. Going from near-zero to 40 percent in one year tells you the locomotion models are eating real-world data fast.
• 0:49
  [Host] Okay, but running on a road isn't the same as folding my laundry. How hard is it really to make a robot run straight for an hour?
• 1:00
  [Guest] Harder than it sounds. Outdoor terrain is uneven, the course has turns, wind gusts shift the center of mass. Running is a controlled fall, and doing it for 13 miles without face-planting requires real-time adaptation that scripted movement can't handle.
• 1:16
  [Host] So the 40 percent that did it without a human in the loop, what changed for them?
• 1:23
  [Guest] [serious] Fleet data. Every robot on last year's course generated sensor logs, whether it was autonomous or remotely piloted. That teleoperation data becomes training signal, because the operator's corrections teach the model what good running looks like on concrete, on turns, on inclines.
• 1:41
  [Host] So the 60 percent that still needed a human operator this year are basically writing the curriculum for next year's autonomous runners?
• 1:50
  [Guest] Exactly. The more units you run, the faster every unit improves. Same flywheel as Tesla collecting driving data from millions of cars, just applied to legs instead of wheels.
• 2:01
  [Host] And this is happening in China specifically because...
• 2:05
  [Guest] Beijing included humanoid robot development as a national strategic priority in its 2026 to 2030 five-year plan. TrendForce published its global humanoid robot shipment forecast in early 2025, projecting about 50,000 units shipping globally this year.
• 2:19
  [Host] And China's share of that?
• 2:21
  [Guest] Morgan Stanley's January 2025 China humanoid industry outlook doubled the country's forecast alone to 28,000 units. TechCrunch reported in February that Chinese companies control roughly 90 percent of global humanoid shipments. When your government designates an industry as strategic and your manufacturers already dominate volume, you can stage a half-marathon as a national showcase.
• 2:43
  [Host] So this race is propaganda?
• 2:45
  [Guest] It's propaganda the same way the space race was propaganda. The rockets still flew. The robot still finished, and the 40 percent autonomy rate is a real engineering result that Western competitors have to contend with.
• 3:00
  [Host] [curious] What does beating the human world record actually mean for the commercial side, though? Nobody's buying a robot to run marathons.
• 3:09
  [Guest] No, but the locomotion and navigation problems overlap heavily with warehouse logistics, construction traversal, elder care in multi-story homes. If your robot handles 13 miles of outdoor terrain autonomously, it can handle a hospital corridor. The race is a stress test disguised as a spectacle.
• 3:27
  [Host] You mentioned the data flywheel. The bottleneck for the whole humanoid industry isn't hardware anymore, right?
• 3:34
  [Guest] Right. Unitree sells a capable robot for $16,000. The real constraint is demonstration data for physical tasks. Every approach to solving it, teleoperation, fleet learning, sensor gloves capturing human movement, is a race to close that data gap.
• 3:49
  [Host] And the country shipping the most robots collects the most training data.
• 3:54
  [Guest] [confident] Which makes the robots better, which means they ship more. AgiBot hit 10,000 cumulative units in March. Unitree is targeting up to 20,000 deliveries this year. Each deployed unit is a data node. China isn't just winning on manufacturing cost. It's winning on the flywheel that turns volume into intelligence.
• 4:15
  [Host] Beijing just standardized that loop by drafting industry-wide engineering standards for humanoid robots. You standardize an industry when you're confident you're leading it.
• 4:24
  [Guest] And that's what Patrick Deglon calls the convergence of the Mind and the Body in un-SCAR-city. The hardware and the intelligence feeding each other in a cycle that's getting harder to catch from behind.
• 4:38
  [Host] For more on the training data bottleneck and how fleet learning accelerates the humanoid industry, head to unscarcity dot ai. The book is on Amazon, the audiobook is on Spotify. This has been Minds, Bodies, and Terawatts.